Multi-Agent Claude Systems
Design and build systems where multiple Claude agents collaborate. Cover orchestrator-worker patterns, agent handoffs, shared state, parallel execution with debate and vote consensus, and specialization so each agent has focused tools and a focused system prompt.
About This Course
Design and build systems where multiple Claude agents collaborate. Cover orchestrator-worker patterns, agent handoffs, shared state, parallel agent execution with debate/vote consensus, and specialization (each agent has focused tools + system prompt). Uses the tool_use loop from CLD-AI-103 as the foundation.
Course Curriculum
10 Lessons
Multi-agent Claude systems: patterns overview
By the end of this lesson you will know the four canonical multi-agent Claude patterns and when to reach for each: orchestrator-worker (a dispatcher decomposes a request and fans out to parallel workers), specialist team (a router picks one focused agent per request), sequential pipeline (agent A's output feeds agent B — draft → edit → format), and debate/vote (multiple agents solve the same problem, an aggregator synthesizes). Sets up L2 through L10 where you build each pattern hands-on.
Orchestrator-worker Claude pattern - Lab Exercises
By the end of this hands-on lab you will have built the orchestrator-worker pattern in Python: a Claude orchestrator uses tool_use with a decompose schema to split a research question into three sub-questions, dispatches all three in parallel via ThreadPoolExecutor to worker Claude instances (each with a focused system prompt), then synthesizes the three answers into a final response. Uses the PythonAI container + Claude API proxy — measures wall-clock speedup from parallelism vs a sequential baseline.
Specialist team pattern
By the end of this lesson you will know how to route customer requests to the right specialist Claude agent using a lightweight classifier — three specialists for billing / technical / feature-request each with their own focused system prompt and narrower tool set — and when the specialist pattern beats a single generalist (lower per-call cost, tighter refusal patterns, sharper responses). Sets up L4 hands-on where you build the router + three specialists end-to-end.
Specialist team routing - Lab Exercises
By the end of this hands-on lab you will have built a Claude router that classifies a support request as billing / technical / feature-request, then dispatches to the matching specialist agent — each specialist configured with a focused system prompt and its own narrower tool set (billing has refund tools, technical has runbook tools, feature-request has ticket-creation tools). Run four end-to-end requests to see routing decisions and specialist responses. Uses the PythonAI container + Claude API proxy.
Sequential pipeline pattern
By the end of this lesson you will know the sequential pipeline pattern — chaining Claude agents where each stage's output becomes the next stage's input (draft → edit → format for a technical blog post; extract → validate → summarize for a document pipeline) — why narrowing each stage's job improves quality over a single mega-prompt, and how to design intermediate schemas so a stage failure is caught before it corrupts downstream stages. Sets up L6 hands-on where you build a three-stage blog-post pipeline.
3-stage sequential pipeline - Lab Exercises
By the end of this hands-on lab you will have built a three-stage sequential Claude pipeline (drafter → editor → formatter) for a technical blog post — each stage has its own focused system prompt (drafter emphasizes technical accuracy, editor prunes fluff, formatter enforces markdown structure) — and observed how narrowing each stage's job produces sharper output than a single mega-prompt trying to do all three. Uses the PythonAI container + Claude API proxy.
Debate + vote pattern
By the end of this lesson you will know the debate+vote pattern — ask N Claude instances (different system prompts, sometimes different models like Sonnet 5 + Opus 5.5 + Fable 5.1) the same high-stakes question in parallel, then have a judge Claude aggregate their answers or synthesize a balanced recommendation — when the extra cost is worth it (product / hiring / regulatory decisions), and why it beats single-shot on tasks where perspective diversity matters. Sets up L8 hands-on.
Debate + synthesize pattern - Lab Exercises
By the end of this hands-on lab you will have built a three-advisor Claude debate: conservative / aggressive / contrarian system prompts run in parallel over the same product decision, then a judge Claude synthesizes a balanced recommendation citing each advisor's key point. Observe how perspective diversity in the advisors surfaces tradeoffs a single Claude call would miss. Uses the PythonAI container + Claude API proxy — measures the added latency + cost vs single-shot.
Multi-agent observability + evaluation
By the end of this lesson you will know how to debug multi-agent Claude systems — structured JSON logging per-agent (trace_id, agent_role, tokens, latency, tool calls), tracing across agent handoffs to reconstruct end-to-end request flow, and evaluating end-to-end quality when N agents each contribute to the final answer (per-agent scoring + an aggregate rubric). Sets up L10 capstone where you wire logging + traces across the four-pattern Orion system.
Multi-agent Orion system capstone - Lab Exercises
By the end of this hands-on capstone you will have wired the full CLD-AI-107 multi-agent Orion assistant: a router picks between a quick specialist (billing lookups), a complex sub-orchestrator (multi-tool technical investigation), and a high-stakes debate agent (policy exceptions) — with structured JSON logs + trace_id propagated across every handoff so you can reconstruct any request. Uses the PythonAI container + Claude API proxy — pulls together every pattern from L2 through L9.